VideoLVR: Latent Visual Reasoning over Grounded Video Evidence
arXiv cs.AI · 28 Oct 2026, 03:41 AEST
VideoLVR introduces a new method for improving video understanding in multimodal large language models. It constructs latent visual evidence from local regions at relevant video moments to enhance reasoning without significantly expanding context.
Key facts
- Chain-of-thought reasoning has substantially advanced video understanding in multimodal large language models (MLLMs).
- Existing approaches maintain visual grounding during reasoning in two main ways: they either refer to evidence using textual coordinates or reinsert cropped regions or frames into the multimodal context.
- To address these limitations, we introduce VideoLVR, which constructs latent visual evidence from local regions at relevant video moments.
- We further extend VideoLVR to reinforcement learning by treating sampled visual tokens as actions alongside text tokens and jointly optimizing the policy using final-answer rewards.
- Experiments on multiple video benchmarks demonstrate its effectiveness, especially on tasks requiring fine-grained visual evidence.
Read aloud
VideoLVR introduces a new method for improving video understanding in multimodal large language models. It constructs latent visual evidence from local regions at relevant video moments to enhance reasoning without significantly expanding context. Chain-of-thought reasoning has substantially advanced video understanding in multimodal large language models MLLMs. Existing approaches maintain visual grounding during reasoning in two main ways: they either refer to evidence using textual coordinates or reinsert cropped regions or frames into the multimodal context. To address these limitations, we introduce VideoLVR, which constructs latent visual evidence from local regions at relevant video moments. We further extend VideoLVR to reinforcement learning by treating sampled visual tokens as actions alongside text tokens and jointly optimizing the policy using final-answer rewards. Experiments on multiple video benchmarks demonstrate its effectiveness, especially on tasks requiring fine-grained visual evidence.
Sources in this cluster
- VideoLVR: Latent Visual Reasoning over Grounded Video Evidence
arXiv cs.AI · 28 Oct 2026, 03:41 AEST